OpenAI pulled Instant Checkout back: exactly the signal to move now

Three dates tell the whole story. September 29, 2025: OpenAI and Stripe release the Agentic Commerce Protocol and Instant Checkout. Etsy products sell inside ChatGPT, the official line promises "over a million Shopify merchants coming soon," and Etsy stock is up 16% that day. February 2026: the real number per Forrester's count, about 30 merchants live (CNBC, March 20, 2026). March 24, 2026: OpenAI announces the redesign: native in-chat checkout is shelved, and the company will "focus our efforts on product discovery," with purchases completed on the merchant's own site.

From a million merchants, to thirty, to a voluntary pivot: the easy read is "AI shopping is retreating." Easy to share, mechanically wrong. To read this pullback correctly, you only have to look hard at one number's denominator.

Walmart's "one third": both paths start in ChatGPT

The decisive testimony came from a Walmart executive: conversion on ChatGPT's native in-chat checkout ran at roughly one third of sending the same user back to walmart.com to check out (CNBC). The structure of that sentence is the point: both paths being compared start from ChatGPT. What the number killed is the act of checking out inside the chat. What the number proved is the road that matters: AI delivering purchase-intent users to the merchant's site. People will take an AI's recommendation; they'd still rather hand their money to a checkout page they know.

Which is the literal content of the official pivot: discovery validated, native checkout not. Rarely has a "retreat" stated the right direction this clearly.

Native checkout didn't clear this round: trust, not technology

• Gartner (May 2026 press release): even in low-risk categories, the ceiling of consumers willing to let AI make the purchase decision outright is about 11%.
• BigCommerce (2026): 67% of US consumers are interested in trying agentic shopping, but most want human approval before the purchase. Help me choose: yes. Buy it for me: not yet.
• Visa's official threat analysis lists a whole new attack surface: agent impersonation, prompt manipulation, synthetic identities, machine-speed fraud.

And there's a gap none of them close on their own: when an agent misreads intent and places the wrong order, who eats the return and the chargeback has no clean answer today. The merchant is asked to trust a buyer it didn't build and can't audit. Three lines, one conclusion: the last mile of agent-completed purchases is blocked on payment trust and liability, not on model capability. Commerce tolerates errors at a rate close to zero. These problems will get solved: AP2's signed mandate chain and the card networks' controllable credentials aim squarely at them. But on infrastructure's clock, not a keynote's.

The same quarter, discovery kept climbing

• Adobe Analytics (direct traffic measurement, not surveys): 2025 holiday-season AI referral traffic to US retail sites up 693% YoY, the fastest-growing channel, with AI-referred visitors converting 31% better and generating 254% more revenue per visit.
• Salesforce (billions-of-transactions Commerce Cloud base): AI influenced about 20% of global online sales that season, roughly $262B. Influenced, not "orders completed by agents."
• Amazon's shopping agent Rufus reached over 300 million customers in 2025, with annualized incremental sales on the order of ten billion dollars (Fortune / PPC Land figures).

Set the two facts side by side: the checkout layer retreated while the discovery layer surged. Not a falsified trend. A corrected path.

The bear case deserves a straight answer

The strongest version of the bear case: there are pointed critiques calling agentic commerce "a collective hallucination," and eMarketer's conservative forecast (~$144B, counting only transactions agents complete outright) is an institutional caution in its own right. What these voices doubt is the agent-buys-for-you half. And that half, OpenAI just spent a quarter and real money testing on the industry's behalf, with the verdict written into its own pivot announcement. Yet nobody's data disputes that AI is becoming the entry point of purchase decisions. Trust, payment liability, error tolerance: every blocker sits on the transaction mile. The discovery mile is already settled by measurement, by Adobe, Salesforce and Amazon. Both readings converge on the same move: whenever the checkout layer finally clears, win discovery first. On that step, both sides' data are pushing you.

For merchants, this is actually the better deal

Traffic arrives from AI; checkout happens on your own store. No platform checkout tax, the customer relationship and data stay fully yours, and conversion lives on pages you can optimize. The whole arrangement has exactly one precondition: the AI has to recommend you first. That is why "win get-found first" is the optimal play under this structure, not a stopgap.

Four things you can do today

1. Demote "integrate in-chat checkout" on your 2026 roadmap. The platform itself changed course; you don't need to be more anxious about it than they are.
2. Spend the effort on being readable by AI: structured product data, crawlable product pages, clean category and shipping-scope fields.
3. Stand up measurement: does your brand appear in each engine's answers; attribute AI referral traffic separately instead of letting it sleep in "Direct."
4. Do your own checkout homework: the traffic returns to your site, so your landing pages and checkout flow are your conversion rate. No platform will do this course for you.

Two traps: don't fall in either direction

• Don't read the pullback as "AI shopping is over." What got shelved is one layer of a five-layer rail stack; the discovery layer set records the same quarter.
• OpenAI retired the initial checkout path while keeping product discovery as its focus. Merchants retain their own checkout, and brands can establish machine-readable product data and continuous measurement before transaction protocols converge.

OpenAI spent a quarter testing "where does checkout live" for the whole industry. The question it left unanswered is the one a merchant can act on today: whom does the AI recommend.